基于改进YOLOv3模型的侧扫声纳沉船目标检测

Shipwreck targets detection in side-scan sonar based on improved YOLOv3 model

针对使用Faster R-CNN模型进行侧扫声纳图像沉船目标检测存在耗时长、效率低以及小目标漏警率高等问题,引入YOLOv3模型并结合侧扫声纳沉船图像数据集特点对模型进行了改进。首先,进行浅层特征融合的多尺度训练,从而增加沉船目标浅层特征在检测中所占比重;然后,使用K-means聚类算法重新设置先验框参数及大小,提高小目标检测精度;最后,采用二分类交叉熵函数改进YOLOv3算法中的损失函数,提高模型的收敛速度和泛化能力。实验结果表明:相比Faster R-CNN模型和传统YOLOv3模型,改进YOLOv3模型的AP值达到89.18%,分别提高了1.46%和0.57%;调和平均值F1达到89.08%,分别提高了2.33%和1.04%;检测图片耗时时间为Faster R-CNN模型的3/50,极大地提高了检测效率。该研究结果验证了改进的YOLOv3模型具有更高的检测精度和效率,对海底沉船搜救具有一定的实际指导意义。

In view of the problems of high time-consuming,low efficiency and high alarm miss rate of small targets for the automatic detection of shipwreck targets in side-scan sonar images by Faster R-CNN model,and based on the traditional YOLOv3 model,the YOLOv3 model and improve the mo-del by combining the characteristics of the side-scan sonar shipwreck data set were introduced.The multi-scale training of shallow feature fusion was carried out to increase the proportion of the shallow features of the shipwreck target in the detection,and the K-means clustering algorithm was used to reset the parameters and size of the prior frame so as to improve the small target recognition and positioning accuracy.Finally,the binary classification cross entropy function was used to improve the loss function of the YOLOv3 algorithm.The experimental results show that compared with the Faster R-CNN model and the traditional YOLOv3 model,the AP value of the improved YOLOv3 model reaches 89.18%,an increase of 1.46%and 0.57%respectively;the harmonic mean F1 reaches 89.08%,which is an increase of 2.33%and 1.04%respectively;and the time taken to detect the image is 3/50 of the Faster R-CNN model,which greatly improves the detection efficiency.Experiments verify that the improved YOLOv3 model has higher detection accuracy and efficiency,and thus it is a guide for searching and rescuing shipwreck undersea.